Teach a call agent with human feedback

How it works

1. 1Bring calls in: webhook, dataset or demo data. RL: episodes and a reward signal from the real environment.
2. 2Find the hot spots: moments are clustered into situations, ranked by size and weakness on a map. RL: prioritising states where the policy is weak.
3. 3Duel: one moment, two hidden and shuffled replies; pick the better, tick why, +10 XP. RL: pairwise human preferences, the signal RLHF trains on.
4. 4Level up: picks become a new version. RL: a policy update, prompt-level now, DPO with LoRA for weights.
5. 5Boss fight: new vs old on the same customers and seeds, with win rate, radar and regressions. RL: held-out evaluation with fixed seeds.
6. 6Crown the champion: only a winner goes live. Connect: the call platform asks the champion for its next reply, and every reply comes back as a live call. RL: the deployed policy and online data collection.

The code

# the inference layer: your call platform asks the champion for its next reply
curl -X POST https://levers9-api.uselevers.com/api/rl/agent/<tenant>/reply \
  -H "X-Levers-Webhook-Key: <secret>" -H "Content-Type: application/json" \
  -d '{"call_id": "CALL-0142", "turns": [{"role": "customer", "text": "..."}]}'
# -> {"reply": "...", "policy": {"name": "v2 · trained"}, "judge": {"total": 76}, "violations": []}

# weights, once a preference version is frozen
cd lab/rl && python train_dpo.py --pack saudi_real_estate --version 3 --base google/gemma-4-E4B-it --gpu A10G